Emerging technology can produce a sustained surge in productivity, as I argued last year in an article on the potential of generative AI (with James Manyika of Google). This is consistent with other estimates, like that of the McKinsey Global Institute.
Generative AI is the first AI with a humanlike capacity to operate in multiple domains and to detect and switch domains based only on conversational prompts. It can talk about inflation, write computer code, do some mathematics—though this is a work in progress. Superhuman pattern recognition ability makes it a powerful digital assistant. Rather than full automation, the better model is machine-human collaboration, or what is sometimes called “augmentation.”
Geoffrey Hinton, a pioneer of modern neural network AI, has a special understanding of the implications. He uses the example of an experienced doctor. While she/he may have treated thousands of patients, medical AI can review and absorb hundreds of thousands. That can make it helpful to the experienced doctor, and even more so for those who are less seasoned. This is consistent with studies of AI applications in other areas, like customer service, where AI digital assistants, trained on past interactions, produced large productivity gains overall and even greater benefits for less experienced agents.
AI is general-purpose technology that has applications across the entire economy, by sector and type of work. This is important, because only general-purpose technologies can produce an economy-wide productivity surge.
AI applications are already being built into personal devices such as phones, thanks in part to advanced semiconductors.
That said, challenges need to be overcome to achieve the potential. One is implementing regulation to prevent misuse of the technology and data. That risk-mitigation regulatory agenda is in process across the globe.
Another is overcoming automation bias, or what Erik Brynjolfsson calls the Turing Trap, the strong tendency to view this technology as full automation and thus a replacement for humans.
This is a common view in the media, business, and policy discussions. The widespread concern about dramatic declines in employment reflects this.
Probably the most important policy issue concerns potential gains. For AI to achieve full economic impact over time, it must be accessible to all sectors of the economy, and to companies large and small. There is little doubt that the massive investments undertaken in industries like technology and finance will have a major impact, but the applications need to get to large employment sectors that tend to lag—like government, health care, construction, and hospitality. Pre-AI studies of digital adoption indicate that this broad diffusion pattern is not guaranteed, that left entirely to market forces divergence is possible or even likely.
Policies for accessibility, diffusion, and skills to help realize the full potential of AI are currently weak in comparison with the intense focus on risk mitigation and misuse. Expanding the former without abandoning the latter is an important element of policy rebalancing. This is not to advocate government’s picking winners or national champions. On the contrary, effective competition policy should be part of the policy portfolio. In addition, part of the focus needs to be on sectors and businesses that may lag in discovery and adoption, small and medium enterprises for instance. And since jobs will change with AI collaborators, retraining and new skills acquisition deserve priority attention.
Challenges to overcome
The potential gains from AI go well beyond countering postpandemic productivity and growth challenges. They are set to impact science and technology research, from biology to physics and materials science, and to play a key role in the energy transition.
Talent, computing power, and rapidly expanding electricity demand are the main barriers to building increasingly powerful generative AI models. Availability of data is not a major constraint. The internet has ample training data. Of course, there is AI that is not in the generative AI category that is powerful and important. AlphaFold, an AI system that predicts three-dimensional structures of proteins, is an example. For this application you need specialized biology data and expert input on how protein folding works.
It is also true that the mega-platforms that are driving the development of generative AI have business models that rely on personal data and very precise targeting. But to train large language models and the like, you do not need personalized and sensitive data.
The systems powerful enough to train models with billions of parameters reside largely in cloud computing systems in the private sector, mostly in the US and China. That, plus the competition for talent, puts science and academia at a disadvantage. Expanding computing infrastructure to a broad community of researchers and innovators is an important policy step needed to democratize building an open community with a good balance between academic and private innovation. Achieving that balance will support widespread diffusion.
Europe risks falling behind the United States and China in developing and applying AI for three reasons. One is the European Union’s relative underfunding of basic research. The second is that it lags in computing power to support research. The third is a failure to fully leverage the large scale of the European economy. With high fixed development costs and relatively low variable costs in digital and AI, scale is a huge advantage in determining return on investment. European capital markets remain fragmented; service market integration is incomplete and hampered by fragmented regulation at the national level. Whether this situation persists or there is a change of direction after the recent European Parliament elections remains to be seen. Two reports to the European Commission—one from Enrico Letta and a forthcoming one from Mario Draghi—advocate elevated investment in digital technology.
China is an AI powerhouse. India, with its strong roots in digital technology, a large and growing internal market, and deep reservoirs of engineering human capital, is likely to be a growing force.
The rest of the emerging market economies may benefit greatly from AI applications, but for the next few years at least, they will be largely consumers of advanced AI technology generated mostly in the US and China.
AI will drive large-scale structural change and disruption for decades. While some will lose jobs via automation or rapid productivity growth, and others will be hired for jobs that are new and created by the technology, it’s the workers in the middle who will be most impacted. Here jobs will not necessarily vanish, but they will change. It will be a disruptive process requiring different skills and a lot of organizational change. Both the private and public sectors have important roles in smoothing the transitions.
With policy support to accelerate diffusion across the entire economy, AI could significantly accelerate economic growth and help productivity growth rebound. And if it relaxes the supply-side constraints that are part of the inflation story, indirectly it could lower real interest rates and the cost of capital over time. In a world that requires trillions of dollars of investment to change the equation for energy efficiency and the green transition, that would help. And in the aging part of the global economy, it would help the younger working population support the older group without undue sacrifice.
Despite the shocks and secular headwinds to growth, we do have the talent and tools to foster growth, inclusion, and sustainability in the global economy—but only if we have the will to use them aggressively but wisely.
